Sparse Communication for Distributed Gradient Descent
This work addresses communication bottlenecks in distributed machine learning, offering practical speed improvements for tasks like image classification and translation, though it appears incremental as it builds on existing sparsity and quantization techniques.
The paper tackled the problem of slow distributed stochastic gradient descent by exchanging sparse updates instead of dense ones, achieving up to 49% speedup on MNIST and 22% on neural machine translation without harming accuracy or BLEU scores.
We make distributed stochastic gradient descent faster by exchanging sparse updates instead of dense updates. Gradient updates are positively skewed as most updates are near zero, so we map the 99% smallest updates (by absolute value) to zero then exchange sparse matrices. This method can be combined with quantization to further improve the compression. We explore different configurations and apply them to neural machine translation and MNIST image classification tasks. Most configurations work on MNIST, whereas different configurations reduce convergence rate on the more complex translation task. Our experiments show that we can achieve up to 49% speed up on MNIST and 22% on NMT without damaging the final accuracy or BLEU.